An Intelligent System for Motor Style Assessment and Training from

Inertial Sensor Data in Intermediate Level Ski Jumping

Heike Brock and Yuji Ohgi

Graduate School of Media and Governance, Keio University, 5322 Endo, 252-0882, Fujisawa-shi, Kanagawa, Japan

Keywords:

Inertial Sensors, Activity Recognition, Motion Analysis, Augmented Motion Feedback, Mobile Motor

Training, Ski Jumping.

Abstract:

In this research we developed a wearable, augmented motion feedback system for ubiquitous training and

motion assessment in mid-level ski jumping. Ski jump motion data captured with a set of inertial sensors were

ﬁrst transformed into meaningful kinematic motion information using an extensive processing system. Next,

derived segment orientations, joint positions and joint angles were used to build and train motion knowledge

on the base of the sport’s common style and judging criteria. This intelligent machine knowledge was then

applied to identify speciﬁc style information within incoming motion data that could be provided to the athlete

as augmented motion feedback via a mobile training application. System validations on a set of test jumping

data showed that style errors could be recognized and displayed well by the implemented system. We therefore

believe the system to be suitable for the provision of kinematic motion feedback that could not be obtained

without an extensive training support environment otherwise. Adding a real-time environment for athlete-

system communication, this could lead to the creation of an ubiquitous training support application in future.

1 INTRODUCTION

The enhancementof motor skill acquisition and motor

learning by additional and augmented performance

feedback is one of the most interesting objectives for

technological support in sports. Particularly impor-

tant for the implementation of future training applica-

tions is the development of techniques that provide

motion information on an easy-to-use basis. This

problem comprises both the use of wearable capture

devices that can be employed under any environmen-

tal condition, and the processing of raw numerical

motion data into intuitive data output. In this work,

we addressed both aspects with the intention of im-

plementing a mobile style assessment and training

support system for intermediate and junior level ski

jumping.

Ski jumping is a very technical sport that is de-

ﬁned by biomechanical and physical laws (e.g. drag

and lift) to a large extend. Erroneous motion execu-

tion and use of aerodynamic forces immediately in-

ﬂuence the performance and can furthermore increase

the risk of fall and injury. However to date, knowl-

edge about ski jumping is mainly based on prac-

tical experience, simulations and wind tunnel mea-

surements (Seo et al., 2004; Marqu´es-Bruna and

Grimshaw, 2009b; Marqu´es-Bruna and Grimshaw,

2009a). Exact kinematic and dynamic properties of

an athlete are quite difﬁcult to measure during the ac-

tual jump. This is due to the sport’s large ﬁeld of mo-

tion activity as well as unstable weather and daylight

conditions. They do not allow for quality data from

conventional video or optical motion capture systems.

As a result, computer based motion analysis meth-

ods for ski jumping are effectively non-existent so far,

and the assessment of a jumping performance is still

a mainly visual task.

By making detailed and accurate ubiquitous mo-

tion information available to coaches and athletes,

new training standards could be set. In ski jumping,

we consider a mobile feedback platform as particu-

larly beneﬁcial for junior and intermediate level ath-

letes. Here, economical and logistical constraints in-

ﬂuence the quality of the general training structures:

for example it is common that many jumps are exe-

cuted within a very short span of time. Consequently,

responsible coaches often observe jumps from one

perspective only (generally the coaches’ stand), while

the assessment of every single jump performance has

to be instantaneous. Internal motor representations in

intermediate level jumpers on the other hand are less

stable than in professional athletes, making additional

Brock, H. and Ohgi, Y.

An Intelligent System for Motor Style Assessment and Training from Inertial Sensor Data in Intermediate Level Ski Jumping.

DOI: 10.5220/0006032901010108

In Proceedings of the 4th International Congress on Sport Sciences Research and Technology Support (icSPORTS 2016), pages 101-108

ISBN: 978-989-758-205-9

101

information on previous motion performances a very

valuable feedback in future. Therefore, the aim of this

work was to develop a wearable framework for appli-

cation in motor style assessment and training of mid-

level ski jumping.

2 DATA COLLECTION

85 ski jumps were collected during summer ski jump

season from four different junior athletes (three ski

jumpers and one Nordic Combined athlete) at a nor-

mal hill with a K-point (indicating the hill’s steep-

est point) of 90 meters. The motion of every athlete

was captured with nine waterproof inertial measure-

ment units of 16 bit quantization rate (Logical Prod-

uct, 2015). The measurement units were of 67x26x8

mm size and 20 g weight and had an internal memory

capacity of 32 MB with an average operation time of

3 hours. Every device contained triads of gyroscopes,

accelerometer and magnetometer for the respective x,

y and z axes. The gyroscopes were speciﬁed with a

full-scale range of ±1500 dps with 0.67 mV/dps sen-

sitivity. Accelerometer speciﬁcation varied in depen-

dence on the placement between either a minimum

full-scale range of ±5 G (body placement) or ±16 G

(ski placement) with 191.7 mV/G sensitivity. Mag-

netic ﬁeld sensors had ±1.2 Ga full-scale range. All

sensor modules were sampled at sf

= 500 Hz.

The sensors were positioned to measure motion of

all limbs and segments relevant for the execution of a

ski jump. The forearms for example were not exclu-

sively captured, since the elbow joints in ski jump-

ing are mostly rigid and the forearms moved in equal

terms with the upper arms. In concrete, the follow-

ing sensor positions were chosen: pelvis and both left

and right thigh, shank, ski close to the tip of the ski

boot and upper arm of the athletes (Figure 1). The

sensors were securely placed directly on the athlete’s

body and ski using adhesive and kinesiology tape be-

fore the beginning of data acquisition.

Since the quality of a ski jump is deﬁned by a

mix of measurable jump properties (length, wind) and

jury-based style assessment, we furthermorecollected

the jump length and style scores of every jump. Both

measures were annotated on paper by an experienced

ski jump judge in real-time and under real judging cir-

cumstances from the judge’s tower. After data acqui-

sition, all score sheets were digitized to serve as an

indicator for the quality of a motion performance in

the following.

Figure 1: Nine sensors were attached to the athlete’s body

and ski to capture the ski jumps.

3 DERIVATION OF BODY

KINEMATICS

Despite the need for quantitative data, only a few

studies addressed the use of inertial sensors for perfor-

mance assessment in ski jumping so far (Ohgi et al.,

2009; Lee et al., 2015; B¨achlin et al., 2010; Chardon-

nens et al., 2012; Chardonnens et al., 2013). This

is mainly due to the sparse raw data: inertial sen-

sors measure acceleration, gravity and magnetic ﬁeld

information, which can generally not display all in-

formation necessary for a complete motion analysis.

Simple characteristics and anomalies of a motion per-

formance can be found from the raw data with statis-

tical measures or spectral Fourier and wavelets ﬁlters.

Body segment orientations and joint positions, as they

can be obtained with other motion capture technolo-

gies, cannot be derived immediately and have to be

computed in a post processing step. Various meth-

ods to determine signiﬁcant information from iner-

tial sensor data have been developed within the last

decade (Madgwick et al., 2011; Euston et al., 2008;

Yun and Bachmann, 2006). For the current study, we

used an independent processing framework that we

specially developed for the determination of full-body

kinematics in ski jumping.

Starting with a raw data input, the processing

framework consisted of the following sequence of

computation steps: (1) determination of initial sensor

orientations with an algorithm based on trigonometric

relations in the ﬁeld measurement vectors (Yun et al.,

2008), (2) compensation of magnetic disturbances to

correctly align every sensor to the global reference

frame, (3) estimation of sensor orientation estimation

by a Complementary Filter (Euston et al., 2008), (4)

sensor-bone alignment to adhere for variations in the

sensor placement and to determine body segment ori-

entations, and (5) computation of relative joint posi-

tions with a forward kinematics approach using man-

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support

102

Figure 2: Schematic overview of the processing framework with initial attitude estimation, compensation of magnetic bias,

sensor orientation estimation, sensor alignment and computation of the output body kinematics.

ually measured segment lengths. By the latter two

steps, the system’s output data in form of kinematic

motion information was provided. Furthermore, we

annotated the time instants of the two characteristic

key-events take-off and landing on the base of the raw

sensor data. Every data capture could then be seg-

mented into its main motion phases in-run, ﬂight and

landing, which yielded an additional data output of

phase-wise body kinematics (Figure 2).

The system accuracy has been successfully tested

in a previous work, where it showed errors of 1 − 3

degrees for the initial orientation estimates and drift

errors of less than 5 degrees over the complete data

capture respectively ski jump.

4 CREATION OF INTELLIGENT

MOTION KNOWLEDGE

After the derivation of accurate and reliable kinematic

motion information, the question was how to trans-

form the obtained data into a meaningful description

(respectively feedback information) of the performed

motion. Similar to the complex knowledge of the

human brain, which acquired the ability to perceive

and understand motion performances during years of

practice and experience, machine learning algorithms

should be utilized to create an artiﬁcial intelligent mo-

tion understanding for this task.

4.1 Quality Measure

As a ﬁrst step, it was necessary to determine a mean-

ingful ground truth measure that could describe all

relevant motion information and that could be used

for the training of the machine knowledge. Criteria

for style assessment are based on the biomechanical

descriptions of the motion and are designed in such

a way that they are universally valid and independent

of an athlete’s anthropometric properties. Good mo-

tion technique is generally also correlating to a higher

ﬂight curve and a longer jump. Therefore, we decided

to use the collected style scores as a ground truth for

this investigation.

According to the ofﬁcial scoring system speci-

ﬁed by the International Ski Federation (FIS, 2013),

marks are not given for good style, but deducted for

faults. A perfect jump is awarded with a maximal

style measure of 20 points per judge, and errors and

deviations from the desired motion style in the motion

phases ﬂight, landing and outrun are ﬁned by distract-

ing points from the maximum score. Faulty behavior

during the ﬂight phase and the landing can be pun-

ished with a maximum point deduction of 5 marks

each and during the outrun with a maximum point de-

duction of 7. On a ﬁne scale, error points are deducted

under the style criteriaC shown in Table 1. They were

used to annotate the collected ski jumps in the present

work and correspond with the segmented jump phase

of ﬂight and landing.

4.2 Machine Learning

The applicability of the mobile training system de-

pended largely on how well the processed inertial sen-

sor data could describe a motion and especially depict

all critical phases and properties that inﬂuenced its

performance quality. Especially important here was

to deﬁne meaningful feature representations that rep-

resented the structure and characteristic of the under-

lying data, and to chose and train a suitable machine

An Intelligent System for Motor Style Assessment and Training from Inertial Sensor Data in Intermediate Level Ski Jumping

103

Table 1: Excerpt from ofﬁcial instructions on the judging of ski jump style. The presented performance errors per motion

phase A (aerial phase) and L (landing phase) and their point deductions will serve as main style reference C in the following.

A Aerial phase errors max. 5.0

1 Insufﬁcient control over body or skis during the formation of the stable and dynamic ﬂight

posture

0.5-2.0

2 Instability (unnecessary motion of the arms, uncontrolled body position, bent knees, not

completely stretched legs)

0.5-1.0

3 Unsymmetrical positioning of the arms 0.5-1.0

4 Unsymmetrical positioning of the legs 0.5-1.0

5 Unsymmetrical positioning or unevenness of the skis 0.5-1.0

L Landing phase errors max. 5.0

1 No Telemark landing at all (feet parallel, single fault) min.2.0

2 No smooth movement/transition from the ﬂight pose to the landing 0.5-1.0

3 Slight Telemark landing, with little bending of the knees only 0.5-1.5

4 Insufﬁcient absorption of the landing impact by the Telemark, or Telemark position is not

maintained until the end of the landing process (instability, too stiff or not fully executed

Telemark position)

0.5-1.5

learning method for the identiﬁcation of style errors

and faulty motion executions.

4.2.1 Motion Features

Research in activity recognition from wearable sen-

sor data has resulted in a wide variety of possible

feature transformations such as statistical raw-signal

based features, event-based features, multilevel fea-

tures derived from clustered statistical occurrences

and kinematic body motion information (Bulling

et al., 2014). Many processing methods used in the

context of sports focus on low-level signal-based fea-

tures and extract information directly from the raw

sensor data (Milosevic and Farella, 2015; Dadashi

et al., 2014; Ghasemzadeh and Jafari, 2011). In

this study, we focused on body-model features de-

scribed by positional and angular data or relations be-

tween body parts and body joints. The reason for this

choice was that such ’motion property time-series’

were closest related to the biomechanical description

of the style criteria C. Besides, temporal execution of

a motion as well as correct timing of key motion pat-

terns are very important aspects of motor skill and the

training of motor sequences.

For the following investigations, we designed a

universal set of body-model features F

that could

also be used in similar applications for any other

movement or sports data (Helten et al., 2011). In

concrete, those features constituted the body kinemat-

ics obtained from the data processing framework (F

and F

), but also included further kinematic motion

information built from angular and positional rela-

tions between certain segments and joints (F

and

) (Table 2).

was computed from the angular information

Table 2: Description of the chosen (time-series) body model

features F

with their feature ID for the respective three

sensor axes.

ID Type Description

φ, θ, ψ Roll, Pitch and Yaw in

the global coordinate

frame

rel

, y

rel

, z

rel

Segment end (joint)

position in the global

x,y,z coordinate frame

∠

s1,s2

Angle between neigh-

boring body segments

s1 and s2

j1, j2

, y

j1, j2

, z

j1, j2

Relative position dif-

ferences of joints j1

and j2

of two spatially related, neighboring body segments

and comprised the joint angles of highest inﬂuence

on the aerodynamic effects of a ski jump: hip, knee,

shoulder, ski elevation, ski opening and arm opening

angle. For F

, the positional relations between right

and left body parts (shoulder,hands, hip, feet, ski tips)

along all three axes were used. Every feature was fur-

thermore rescaled to the interval [0, 1]. This rescaling

standardized the feature range and made the features

invariant to anthropometric differences (e.g. different

body segment lengths) between athletes.

4.2.2 Error Classiﬁcation Method

Conforming to the general style assessment of ski

jumping, the basic idea for the creation of motion

feedback information was to classify an input jump as

either error jump (EJ) or non-error jump (NJ) with re-

spect to all nine chosen style criteria. To address this

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support

104

problem, we chose a straight-forward implementation

of the binary support vector machine (SVM). The

principal idea then was to determine commonalities

between two or more motion performances (meaning

motion feature streams) of the same group EJ or NJ.

To handle temporal variations in the time-serial

features, we added a data transformation based on

a weighted-sum singular value decomposition be-

fore the main computation of the SVM (Li et al.,

2005). This strategy can represent temporal informa-

tion within a lower dimension as the concatenation

of the projected ﬁrst singular vector to the ﬁrst prin-

cipal component and the normalized singular value

vector of the motion matrix. Closeness between two

data streams is then registered if their resulting re-

duced feature vectors are of similar value, and differ-

ence if their vector elements are dissimilar. After fea-

ture transformation, the single feature vectors could

be concatenated and used as input data for the SVM.

For the validation and evaluation of a machine

learning system, it is common to have at least two

different data sets: one used to learn the system and

one used to test the trained system. To produce such

a data base split within the 85 ski jump captures, we

made use of the A and L phase-wise ground truth style

annotations. For every C, the numbers of EJ and NJ

jumps were determined and half of each jumps and

their respective phase-annotated feature streams ran-

domly assigned to the training data base. All remain-

ing jumps were assigned to the testing database. To

become more robust against random inﬂuences of the

splitting process into training and testing database,

we chose to use a k-fold cross validation (CV) with

k= 2. This means that the classiﬁcation was per-

formed twice, whereas all data was used once for

training and once for testing. To improve results, we

furthermore added an internal 2-fold CV cycle for the

training of the model parameters of the SVM to the

main cycle, leading to a nested k-fold CV (Figure 3).

Figure 3: Schematic overview of the implemented nested

cross-validation for the learning and validation of the intel-

ligent machine knowledge.

5 RESULTS

As a measure for the classiﬁcation accuracy we com-

puted the precision and recall of the error annotation

in the testing data, whereas the precision and recall

values of both k-fold validation steps were averaged

to yield a ﬁnal output. Under the given problem, pre-

cision (P) was deﬁned as the number of correctly clas-

siﬁed errors n

dividedby the number of all classiﬁed

errors n

f p

. Recall (R) was deﬁned as the number

of correctly classiﬁed errors n

divided by the num-

ber of all elements that are actual errors n

+ n

for

every C:

P =

+ n

f p

, R =

+ n

. (1)

Precision could hence be thought of as a measure of

the classiﬁcation’s exactness, and recall as a measure

of the classiﬁcation’s completeness.

In general, a low precision can indicate a large

number of false positives n

f p

, and a low recall many

false negatives n

. Ideally, both measures should

be close to 1 to show a good error recognition accu-

racy. Results of the nested CV showed that the imple-

mented system was capable to retrieve errors of good

accuracy in all style criteria C, and of high accuracy

in most C, by either high P or R values (Figure 4).

A1 A2 A3 A4 A5 L1 L2 L3 L4

0.2

0.4

0.6

0.8

Error Recognition Precision

A1 A2 A3 A4 A5 L1 L2 L3 L4

0.2

0.4

0.6

0.8

Error Recognition Recall

Figure 4: Precision and recall for the error recognition along

all style criteria C. In a perfect retrieval, both metrics would

be 1.

To obtain a combined accuracy measure repre-

senting all relevantclassiﬁcation statistics, we further-

more computed the normalized confusion matrices of

all C (Figure 5). They contained the retrieval parame-

ters n

, n

f p

, n

and the number of true negatives n

whereas a good classiﬁcation was depicted by high

An Intelligent System for Motor Style Assessment and Training from Inertial Sensor Data in Intermediate Level Ski Jumping

105

Actual

Predicted

Actual

Predicted

Actual

Predicted

Actual

Predicted

Actual

Predicted

Actual

Predicted

Actual

Predicted

Actual

Predicted

Actual

Predicted

Figure 5: Confusion matrices for the error recognition along

all style criteria C. The darker the color along the diagonal

axis (ﬁrst and fourth quadrant), the better the classiﬁcation.

values along the diagonal axis (ﬁrst and fourth quad-

rant) of the matrix:

f p

In the present visualization, high values were denoted

as black and low values (with a min value 0) as white.

Looking at the classiﬁcation accuracy of every C,

we realized that the features of precise error recogni-

tion were those features with clearly deﬁned motion

properties (e.g. A2, A3 and A5). Features of less ac-

curate error recognition on the other hand were gener-

ally less speciﬁc with respect to their deﬁnition in the

judging criteria (e.g. L2, L3, L4). One explanation

here could be that such error annotations got inter-

fused within the process of ground truth data acqui-

sition, since their description was of similar seman-

tic content. Further improvement of the results might

therefore be achievedby a larger collection of training

data for the classiﬁer and more robust groundtruth an-

notations. This could be achieved by simultaneously

collecting style evaluations from several judges that

would then be averaged and reduce the inﬂuence of

misperception.

6 USE IN TRAINING SYSTEM

Knowing the accuracy of the system, the error recog-

nition method should be used for the provision of

motion feedback and style information to the athlete

in the last step. The idea here was to implement a

graphical user interface that can communicate with

the athlete to give directed feedback on the motion

(Figure 6).

In concrete, the design of the athlete-system com-

munication should be as follows. First, incoming

Figure 6: Sample implementation of a graphical user inter-

face for the provision of directed feedback to the athlete.

sensor data of a current motion performance is re-

ceived, processed and classiﬁed under the style cri-

teria C. Once the basic system computation is done,

the athlete can ask for speciﬁc information on mo-

tion parts or motion properties by sending retrieval

requests. Next, the respective information will be re-

trieved and delivered to the user.

Here, it is important to note that search criteria and

keywords for communication with the training system

were held general and intuitive by pre-deﬁned search

queries. Internally, those search queries were associ-

ated to one of the nine style criteria for information

retrieval. A possible query in the user front end could

for example be whether the arms have been held par-

allel during ﬂight. In the back end this information

would be labeled under the criteria A3, and the respec-

tive error recognition result for A3 could therefore be

used to display an either positive (in case of NJ) or

negative (in case of EJ) output feedback (Figure 7).

Figure 7: Sample overlook on the dialog between athlete

and training system.

7 CONCLUSION AND OUTLOOK

In this work, we presented a novel approach for the

provision of automatic motion feedback on the base

of biomechanical style criteria for intermediate level

ski jumping. First, motion performances were cap-

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support

106

tured using nine inertial sensors. The inertial sensor

data were then processed, so that relevant motion in-

formation and body kinematics were obtained. Next,

a fundamental intelligent motion understanding de-

ﬁned by the guidelines for ski jump style assessment

was built from the augmented motion data. This ma-

chine knowledge could then be used to provide mo-

tion feedback information to the athlete by simple

pre-annotated search queries.

Validation of the underlying system methods

showed that the system was capable to identify style

differences and errors well. To enable a more speciﬁc

training system for individual athletes, it might next

be reasonable to use different quality measures inde-

pendent of universal style criteria. This could for ex-

ample mean to include numerical parameters known

to inﬂuence a ski jump performance such as the body

forward angle or the ski attack angle. Considering

that the ideal ﬂight style varies for every athlete in

dependence on his or her individual anthropometrics

and motor skills, it could furthermore be useful to

build individual motion knowledge for every athlete.

Data could then also be used to monitor the progres-

sion of skill over time. However, this would require a

large data base of jumps per athlete before a meaning-

ful motion knowledge could be created – something

which is difﬁcult to organize in practice.

The two biggest issues the system currently has to

face are the provision of real-time feedback, as well

as the correct handling and attachment of the motion

sensors required for a future independent system use

by athletes. Whereas the former can be addressed by

the establishment of a wireless data network for data

transmission at the ski jump hill, the latter is subject

to the user. Consequently, possible sources of error

should be held as small as possible. With the ongo-

ing process of hardware enhancement, sensors would

ideally be smaller and easier to use in future, such as

for example by inclusion within the jump suit. To im-

prove the system and verify its effect and usability, it

is furthermore sensible to test the system under real

conditions in near future.

All in all, we believe that the developedsystem is a

very promising and powerful approach to the question

of future motor training systems. We have shown that

it is possible to provide and directly deliver motion

information by learned machine knowledge. Espe-

cially in intermediate level sports – where the internal

representation of a motor task is unstable and coach-

ing feedback might be unavailable or incomplete –

augmented motion information acquired by means of

such mobile platform could considerably support cor-

rect motor skill acquisition. Ideally, it could enhance

the training environment, and hence contribute to im-

proved motor understanding, motor skill acquisition

and safety.

REFERENCES

B¨achlin, M., Kusserow, M., Tr¨oster, G., and Gubelmann, H.

(2010). Ski jump analysis of an olympic champion

with wearable acceleration sensors. In 2010 Inter-

national Symposium on Wearable Computers (ISWC),

pages 1–2. IEEE.

Bulling, A., Blanke, U., and Schiele, B. (2014). A tuto-

rial on human activity recognition using body-worn

inertial sensors. ACM Computer Survey, 46(3):33:1–

33:33.

Chardonnens, J., Favre, J., Cuendet, F., Gremion, G., and

Aminian, K. (2013). A system to measure the kine-

matics during the entire ski jump sequence using iner-

tial sensors. Journal of Biomechanics, 46(1):56–62.

Chardonnens, J., Favre, J., Le Callennec, B., Cuendet, F.,

Gremion, G., and Aminian, K. (2012). Automatic

measurement of key ski jumping phases and tempo-

ral events with a wearable system. Journal of Sports

Sciences, 30(1):53–61.

Dadashi, F., Millet, G., and Aminian, K. (2014). Estima-

tion of front-crawl energy expenditure using wearable

inertial measurement units. IEEE Sensors Journal,

14(4):1020–1027.

Euston, M., Coote, P., Mahony, R., Kim, J., and Hamel, T.

(2008). A complementary ﬁlter for attitude estimation

of a ﬁxed-wing uav. In IEEE/RSJ International Con-

ference on Intelligent Robots and Systems, 2008. IROS

2008., pages 340–345. IEEE.

FIS (2013). The international ski competition rules (ICR).

Book III. Ski jumping.

Ghasemzadeh, H. and Jafari, R. (2011). Coordination

analysis of human movements with body sensor net-

works: A signal processing model to evaluate baseball

swings. IEEE Sensors Journal, 11(3):603–610.

Helten, T., Brock, H., M¨uller, M., and Seidel, H.-P. (2011).

Classiﬁcation of trampoline jumps using inertial sen-

sors. Sports Engineering, 14(2-4):155–164.

Lee, T. J., Zihajehzadeh, S., Loh, D., Hoskinson, R., and

Park, E. J. (2015). Automatic jump detection in ski-

ing/snowboarding using head-mounted mems inertial

and pressure sensors. Proceedings of the Institution

of Mechanical Engineers, Part P: Journal of Sports

Engineering and Technology, 229(4):278–287.

Li, C., Khan, L., and Prabhakaran, B. (2005). Real-

time classiﬁcation of variable length multi-attribute

motions. Knowledge and Information Systems,

10(2):163–183.

Logical Product (2015). Sports sensing 9-axial water-

proof inertial sensor (ss-ws1215/ss-ws1216). Ac-

cessed 2015-10-17.

Madgwick, S. O., Harrison, A. J., and Vaidyanathan, R.

(2011). Estimation of imu and marg orientation us-

ing a gradient descent algorithm. In 2011 IEEE

An Intelligent System for Motor Style Assessment and Training from Inertial Sensor Data in Intermediate Level Ski Jumping

107

International Conference on Rehabilitation Robotics

(ICORR), pages 1–7. IEEE.

Marqu´es-Bruna, P. and Grimshaw, P. (2009a). Mechanics of

ﬂight in ski jumping: Aerodynamic stability in pitch.

Sports Technology, 2(1-2):24–31.

Marqu´es-Bruna, P. and Grimshaw, P. (2009b). Mechanics

of ﬂight in ski jumping: aerodynamic stability in roll

and yaw. Sports Technology, 2(3-4):111–120.

Milosevic, B. and Farella, E. (2015). Wearable inertial sen-

sor for jump performance analysis. In Proceedings of

the 2015 Workshop on Wearable Systems and Appli-

cations, WearSys ’15, pages 15–20, New York, NY,

USA. ACM.

Ohgi, Y., Hirai, N., Murakami, M., and Seo, K. (2009).

Aerodynamic study of ski jumping ﬂight based on in-

ertia sensors (171). In The Engineering of Sport 7,

pages 157–164. Springer.

Seo, K., Murakami, M., and Yoshida, K. (2004). Optimal

ﬂight technique for v-style ski jumping. Sports Engi-

neering, 7(2):97–103.

Yun, X. and Bachmann, E. R. (2006). Design, implemen-

tation, and experimental results of a quaternion-based

kalman ﬁlter for human body motion tracking. IEEE

Transactions on Robotics, 22(6):1216–1227.

Yun, X., Bachmann, E. R., and McGhee, R. B. (2008).

A simpliﬁed quaternion-based algorithm for orienta-

tion estimation from earth gravity and magnetic ﬁeld

measurements. IEEE Transactions on Instrumentation

and Measurement, 57(3):638–650.

icSPORTS 2016 - 4th International Congress on Sport Sciences Research and Technology Support

108